humans& Released Persimmon, a 550B User Simulator That Fools LLM Judges About 20% of the Time in Group Chats
humans&·low signal
Persimmon is built on Nemotron 3 Ultra 550B and trained with RL against an adaptive discriminator to simulate real people in multi-user conversations. In a multi-user Turing test it fooled LLM judges 21.1% of the time on TIDES, 18.6% on internal workspace chats and 19.8% on TutorMoments, while frontier assistants stayed under 3%. For agent builders, the use case is a realistic simulated user for testing multi-party agents before they meet real people. Access is a limited playground and API.